Using Part-of-Speech N-grams for
Sensitive-Text Classification
Graham McDonald
Craig Macdonald
Iadh Ounis
University of Glasgow
Scotland, UK
g.mcdonald.1@
research.gla.ac.uk
University of Glasgow
Scotland, UK
craig.macdonald@
glasgow.ac.uk
University of Glasgow
Scotland, UK
iadh.ounis@
glasgow.ac.uk
ABSTRACT
Freedom of Information legislations in many western democracies, including the United Kingdom (UK) and the United
States of America (USA), state that citizens have typically
the right to access government documents. However, certain
sensitive information is exempt from release into the public domain. For example, in the UK, FOIA Exemption 27
(International Relations) excludes the release of Information that might damage the interests of the UK abroad.
Therefore, the process of reviewing government documents
for sensitivity is essential to determine if a document must
be redacted before it is archived, or closed until the information is no longer sensitive. With the increased volume of
digital government documents in recent years, there is a need
for new tools to assist the digital sensitivity review process.
Therefore, in this paper we propose an automatic approach
for identifying sensitive text in documents by measuring the
amount of sensitivity in sequences of text. Using government
documents reviewed by trained sensitivity reviewers, we focus on an aspect of FOIA Exemption 27 which can have a
major impact on international relations, namely information
supplied in confidence. We show that our approach leads to
markedly increased recall of sensitive text, while achieving
a very high level of precision, when compared to a baseline
that has been shown to be effective at identifying sensitive
text in other domains.
1.
INTRODUCTION
Freedom of Information (FOI) laws exist in many countries around the world, including the United Kingdom (UK)1
and the United States of America (USA)2 . FOI states that
government documents should be open to the public. However, many government documents contain sensitive information, such as personal or confidential information. Therefore, FOI laws make provisions that exempt sensitive information from being open. To avoid the accidental release
1
2
http://www.legislation.gov.uk/ukpga/2000/36/contents
http://www.foia.gov
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than the
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from Permissions@acm.org.
ICTIR’15, September 27–30, Northampton, MA, USA.
Copyright is held by the owner/author(s). Publication rights licensed to ACM.
ACM 978-1-4503-3833-2/15/09 ...$15.00.
DOI: http://dx.doi.org/10.1145/2808194.2809496 .
of sensitive information, it is essential that all government
documents are sensitivity reviewed prior to release.
Sensitivity reviewers are required to identify any sequences
of sensitive text in a document, for example individual terms,
sentences, paragraphs or the full content of the document, so
that the sensitive text can be redacted or the document can
be closed. However, the recent increase in volume of digital
government documents means that the traditional manual
review process is not feasible for digital government documents. Therefore, the UK and the USA governments have
recently recognised the need for automatic tools to assist the
sensitivity review process [2, 3].
In this work, we address the problem of automatically
identifying sensitive information in government documents
by directly classifying the text within documents. We present
an approach that uses POS n-grams to measure the amount
of sensitivity contained within a sequence of text, i.e. its
sensitivity load. Moreover, we show how our approach can
be used to deploy an effective sensitivity classifier that classifies sensitivity at the term-level.
We propose to perform classification at the term-level because correctly identifying partial sensitive sequences will be
beneficial to a sensitivity reviewer. For example, by drawing
a reviewer’s attention to the positions of sensitivities within
a document it will likely reduce the time taken to review the
document.
We initially focus on an aspect of FOIA Exemption 27 International Relations, namely “information supplied in confidence”, since this sensitivity has a clear potential to cause
damage to international relations if inadvertently released
into the public domain. We show that our approach can
markedly improve recall of in confidence sensitivities compared to a baseline approach that has been shown to achieve
high levels of recall for sensitive text in other domains [11].
The remainder of this paper is as follows. We present
related work in Section 2. We present our approach in Section 3, before presenting two classification methods, in Section 4 and Section 5, that implement our approach. In Section 6 we present a baseline approach that we compare our
results against. We present our experimental setup in Section 7 and our results in Section 8. Finally, we present our
conclusions and future work in Section 9.
2.
RELATED WORK
Gollins et al. [5] provided an overview of the challenges
presented by digital sensitivity review and how these challenges can be addressed by Information Retrieval (IR) techniques. In that work, they noted that although the type of
sensitivity can be identified, for example personal information, the existence of many sensitivities rely not only on the
terms or entities in the document but also on the context of
the information, i.e. what is said about an entity, how it is
said and who said it.
In [9], we presented an initial approach to address the issue
of context dependent sensitivity, highlighted by Gollins et al.,
by deploying a text classification approach with additional
features such as the entities in the document, a country risk
score and a subjective sentences count to identify documents
that contained personal information and international relations sensitivities. Differently from that work, in this paper
we investigate automatically identifying sequences of sensitive text, within documents, that relates to information
supplied in confidence.
Outwith the field of sensitivity review, most work on automatically identifying sensitive text in documents is in the
field of document sanitization [1, 4]. Document sanitization
tries to automatically mask personal information, or information that could reveal the identity of the person the document is about. A popular approach for this is Named Entity
Recognition (NER). NER typically identifies person, location and organisation entities in text and NER approaches
to document sanitization typically assume that all named
entities in a document are sensitive. However, in this work
we try to automatically identify information that has been
supplied in confidence and, as such, we need a more general
solution that can identify sensitivities in what is said about,
or by, an entity.
Sánchez et al. [11] presented an approach to sensitive
text identification that is more general than NER. They assumed that sensitive text is likely to be more specific than
non-sensitive text, and used the Information Content (IC)
of noun phrases as a measure of how sensitive the phrase is.
Sánchez et al. focused on identifying personal information
sensitivities. However, they also identified textual phrases
that are potentially confidential. Therefore, their work is
more closely aligned to identifying FOIA Exemption 27 sensitivities than NER approaches. Moreover, identifying all
potential sensitivities in government documents is the first
stage of the sensitivity review process and Sánchez et al.
found that approach achieved higher recall for sensitive information when compared to NER. For these reasons, we use
that approach as a baseline for comparing our work against
and we describe our implementation of it in Section 6.
3.
IDENTIFYING SENSITIVITY LOADED
PART-OF-SPEECH N-GRAMS
The approach we present in this paper uses the sensitivity
load of POS n-grams to identify sequences of sensitive text in
documents. The approach is inspired by Lioma and Ounis [8]
who showed that the distribution of POS n-grams in a corpus
can indicate the amount of information they contain. More
specifically, Lioma and Ounis showed that high frequency
POS n-grams are typically content rich and removing content poor POS n-grams from search engine queries results in
an improved overall retrieval performance. Differently from
their work, we use the content load of POS n-grams to try
to measure the sensitivity load of specific sequences of text.
Our intuition is that certain POS sequences might be more
frequent in specific sensitivities. For example, sensitivities
relating to information supplied in confidence would likely
contain variations of the sequence noun - verb - pronoun, indicating that someone has supplied information to someone
else. Our approach uses the distribution of POS n-grams to
try to identify sequences that are specific to this sensitivity.
To do this, we first represent documents by the POS
n-grams they contain. For example, the sequence “The envoy will report on Tuesday” results in the POS tags “DT
NN MD VB IN NN”. Representing this sequence as POS
3-grams results in the following “DTNNMD NNMDVB MDVBIN VBINNN”
Having represented the documents by their POS n-grams,
we then use a probabilistic method to measure the sensitivity load of a POS n-gram. More specifically, following the
work of Li et al. [7], we first construct a 2-way contingency
table as shown in Table 1, where pos, s is the number of
documents in which the POS n-gram appears in sensitive
text, pos, ¬s is the number of documents in which the POS
n-gram appears in non-sensitive text, ¬pos, s is the number
of documents that do not contain the POS n-gram in sensitive text and ¬pos, ¬s is the number of documents that do
not contain the POS n-gram in non-sensitive text.
Table 1: 2-way contingency table used to calculate
the Chi-square statistic of a part-of-speech n-gram.
sensitive non-sensitive
Containing POS
pos, s
pos, ¬s
Not Containing POS
¬pos, s
¬pos, ¬s
Having constructed the contingency table, we use the ChiSquare test of independence to measure the degree of dependency between a POS n-gram and sensitive text. The
2
Chi-square score for a POS n-gram, (Xpos
) is calculated as
follows,
2
Xpos
=
Nd (p(pos,s)p(¬pos,¬s)−p(pos,¬s)p(¬pos,s))2
p(pos)p(¬pos)p(¬s)p(s)
(1)
where p(pos, s) is the probability that the sensitive text
of a document contains the POS n-gram, p(pos, ¬s) is the
probability that the non-sensitive text of a document contains the POS n-gram, p(¬pos, s) is the probability that the
sensitive text of a document does not contain the POS ngram, p(¬pos, ¬s) is the probability that the non-sensitive
text of a document does not contain the POS n-gram, p(pos)
is the probability that a document contains the POS n-gram,
p(¬pos) is the probability that a document does not contain
the POS n-gram, p(s) is the probability that text in the
collection is sensitive, p(¬s) is the probability that text in
the collection is not sensitive and Nd is the total number of
documents in the collection.
The Chi-square test of independence measures how much
the observed frequency of a POS n-gram diverges from its
expected frequency within a corpus. If a POS n-gram’s ChiSquare score is greater than the Chi-Square distribution’s
critical value for a 95% confidence level with one degree of
freedom, then we assume that the distribution of the POS
n-gram in the corpus is related to sensitive text. We refer
to these n-grams as being sensitivity loaded.
By applying our approach, the identified sensitivity loaded
POS n-grams can then be used as features in any method
for automatic sensitive-text classification. In Section 4 and
Section 5, we present two such methods that integrate our
approach.
4.
SENSITIVITY LOAD FILTERING
Confidential information in documents is rare. Indeed,
95% of terms in our collection are in fact part of sequences
that the reviewers did not believe to be sensitive. Moreover,
many documents adhere to template structures in which certain sections of a document are particularly unlikely to contain confidential information, for example the header section
of an email stating the sender, recipients and a subject title.
Therefore, we can use the sensitivity loaded POS n-grams
identified by our approach, presented in Section 3, to filter
out sequences of text that are most likely to contain nonsensitive information.
To do this, we count the number of times a sensitivity
loaded POS n-gram appears in sensitive and non-sensitive
text and use the ratio of sensitive and non-sensitive occurrences to identify POS n-grams that are representative of
non-sensitive sequences. Equation 2 shows how we calculate the ratio, lRatiopos,s , where Sensitive correspondes to
when lRatiopos,s > 1 and Non-Sensitive corresponds to when
lRatiopos,s < 1.
lRatiopos,s =
p(pos,s)p(¬pos,¬s)−p(pos,¬s)p(¬pos,s)
p(pos)p(s)
+ 1 (2)
To classify terms in a given sequence of text k, we represent
k by its POS n-grams kpos . Then, for any kpos in the set of
previously identified non-sensitive POS n-grams, each term
t in kpos is classified as being Non-sensitive, all other terms
are classified as sensitive. We refer to this method as the
F iltering method.
5.
SENSITIVITY LOAD SEQUENCES
The identified sensitivity loaded POS n-grams can be used
to generate features for a Conditional Random Fields (CRF)
sequence tagger [6]. A CRF is a probabilistic framework
that models the conditional distribution p(y|x) in sequential
data, where x is a sequence of observations, an observation
is a term with a set of features that describe the term and
y is a sequence of class labels.
To deploy the CRF method we use two term features. The
first feature we use is the term’s POS tag. Additionally, we
use the sensitivity loaded POS n-grams identified by our
approach to generate a feature tag that indicates whether
the term is part of a sequence that maps to a sensitivity
loaded POS n-gram.
To generate the feature tag, we look at a sequence of n
(POS tagged) terms at a time and check if the sequence
maps to a sensitivity loaded POS n-gram. If a mapping
exists, we tag each term in the sequence. We then move the
sliding window by one term to the next sequence and repeat
the process.
To illustrate this, we return to our example from Section 3.
Recall that the sequence “The envoy will report on Tuesday”
is represented by the POS 3-grams “DTNNMD NNMDVB
MDVBIN VBINNN”. If our approach identifies “NNMDVB”
Figure 1: Illustration showing the POS tag and generated sensitivity loaded feature tags for a sequence.
as the only sensitivity loaded n-gram, this would result in
the sequence being tagged as shown in Figure 1.
When the learned CRF model is deployed, it predicts class
labels for each term in a sequence based on its previous
observations. We refer to this method as CRF +P OS +T AG.
We also present the results for the CRF method without the
sensitivity loaded feature tag, CRF +P OS, and the simple
CRF (i.e. using just the term) as CRF .
6.
INFORMATION CONTENT
As previously mentioned in Section 2, we compare our approach against a baseline from the literature that uses the
Information Content (IC) of a noun phrase as a measure of
the phrase’s sensitivity [11]. A noun phrase is a sequence
of terms that has a noun or indefinite pronoun at the head
of the phrase. IC measures the amount of information provided by the sequence of terms, within the context of a background corpus. More specific term sequences are considered
as having a higher likelihood of being sensitive.
To calculate the IC of noun phrases in a document, the
document is first parsed to extract its syntactic structure.
Noun phrases are then extracted from the resulting syntax tree and submitted to a Web search engine as a query.
The IC of the noun phrase is calculated using the number
of returned results as an indication of the phrase’s specificity. Formally, the IC of a noun phrase, np, is computed
res(np)
, where res(np)
as IC(np) = − log2 p(np) = − log2 totalpages
is the number of returned search results and totalpages is
the number of sites indexed by the search engine. For our
experiments, each term within a noun phrases with an IC
score greater than an empirically defined threshold, β, is
classified as being sensitive. All other terms are classified as
non-sensitive. This baseline is referred to as Inf Content.
7.
EXPERIMENTAL SETUP
Collection: The collection consists of government documents with “information supplied in confidence” sensitivities. The documents have been sensitivity reviewed by
trained sensitivity reviewers. Reviewers were asked to annotate the sensitive sequences within the documents, therefore
the documents have term-level class labels i.e. each term
within a sensitive sequence is labelled sensitive and all terms
that were not annotated are labelled non-sensitive. There
are a total of 231893 terms in the collection with 10838 sensitive terms and 221055 non-sensitive terms, in a set of 143
documents. For our experiments, we split the collection at
the document level to retain the context of the terms and
perform a 5-fold Cross Validation.
Baselines: We compare our approach against the IC
baseline presented in Section 6. We use Open NLP3 to extract noun phrases from documents and for calculating the
IC score of noun phrases we use the Bing search engine4
and set the total number of sites indexed, totalpages, to
3.5 billion. For our experiments we test IC threshold values
of β = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 20, 30, 40, 50, 75, 100}. In future work, we intend to investigate learning the β parameter.
We also report the results of a simple classifier that classifies all terms as the majority class, referred to as All NonSensitive. Conversely, we also report a classifier that classifies all terms as sensitive, referred to as All Sensitive.
3
4
https://opennlp.apache.org
http://www.bing.com/
Sensitivity Load: For identifying sensitivity loaded POS
n-grams in Section 3, we use the TreeTagger5 for POS tagging and, following Lioma and Ounis [8], we use a reduced
set of 15 POS tags. We calculate the Chi-Square statistic
on the training data for n = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10}. In
future work, we intend learning the optimum value for n.
Classification: For Sensitivity Load Filtering (Section 4)
and Sensitivity Load Sequences (Section 5), we test the
methods for each value of n = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10}. For
the Sensitivity Load Sequences classification method, we use
the Mallet6 linear chain CRF tagger.
Metrics: We report balanced accuracy (BAC) due to the
imbalanced nature of our collection. Moreover, sensitivity
review is a recall-oriented task since reviewers must identify
all sensitivities to avoid inadvertent release. Therefore, we
report the F2 measure that provides a weighted average of
precision and recall where recall is attributed more importance and, therefore, a greater weight. We also report the
standard accuracy, precision and recall metrics.
8.
RESULTS
Table 2 shows the best achieved performance for the Filtering method (F iltering), the CRF method using the POS
tag and sensitivity load as features (CRF +P OS +T AG), the
CRF method with the POS tag only (CRF +P OS) and the
simple CRF (CRF ). Table 2 also shows the best performance for the Information Content baseline (Inf Content)
and the All Sensitive and All Non-Sensitive classifiers scores.
The first thing we note from Table 2 is that the CRF classification metod using the sensitivity loaded POS n-grams
identified by our approach outperforms the IC baseline for all
metrics. The CRF method using sensitivity loaded 10-grams
(CRF +P OS +T AGn=10 ) achieves 0.4573 recall. Importantly,
the method also achieves 0.9992 precision and, therefore,
achieves a balanced accuracy of 0.7282. On closer inspection of the results we found that the CRF method with this
setting identified 99% of sensitive text in 67% of the documents and, therefore, we believe this method would be useful in assisting the sensitivity review process. Moreover, we
note that the CRF method without the sensitivity loaded
POS n-gram feature (CRF +P OS) correctly identified less
than 5% of the sensitive text. This further demonstrates
the effectiveness of our approach.
The next point we note from Table 2 is that the Filtering
approach also achieves its highest recall (0.9129) when n is
set to 10. However, the approach achieves a low precision
score (0.0563) resulting in an F2 score of (0.2257). This low
level of precision shows that, on this collection, this method
is markedly over-predicting sensitivity.
We also see from Table 2 that the CRF and the Filtering
methods achieve their best recall scores using POS 10-grams.
Finally, we note that the IC baseline, Inf Contentic>7 ,
achieved 0.0116 recall and 0.2564 precision. The IC baseline
identifies syntacticly complex phrases and specific terms as
being sensitive. However, not all complex or specific phrases
are confidential and, moreover, not all sensitivities are noun
phrases. We also note that the classification methods that
use our approach for identifying sensitive sequences of text
achieve markedly better recall, and notably better balanced
accuracy, than the IC baseline.
5
6
http://www.cis.uni-muenchen.de/ schmid/tools/TreeTagger
http://mallet.cs.umass.edu
Table 2: Results for the Filtering and CRF methods.
The table also shows the performance of the All Sensitive,
All Non-Sensitive and Information Content baselines.
all Non-Sensitive
all Sensitive
Inf Contentic>7
Inf Contentic>8
Inf Contentic>9
Inf Contentic>10
Inf Contentic>20
F ilteringn=7
F ilteringn=8
F ilteringn=9
F ilteringn=10
CRF
CRF +P OS
CRF +P OS +T AGn=7
CRF +P OS +T AGn=8
CRF +P OS +T AGn=9
CRF +P OS +T AGn=10
9.
accuracy
0.9533
0.0467
0.9457
0.9457
0.9457
0.9458
0.9459
0.3763
0.3360
0.2616
0.1815
0.9226
0.9221
0.9600
0.9637
0.9657
0.9712
balanced Acc
0.0000
0.5234
0.1340
0.1245
0.1188
0.1279
0.0983
0.3805
0.4007
0.4434
0.4846
0.0390
0.0647
0.5702
0.6264
0.6608
0.7282
precision
0.0000
0.0467
0.2564
0.2388
0.2281
0.2466
0.1911
0.0579
0.0573
0.0570
0.0563
0.0179
0.0856
0.8310
0.8932
0.9454
0.9992
recall
0.0000
1.0000
0.0116
0.0103
0.0094
0.0093
0.0055
0.7032
0.7441
0.8298
0.9129
0.0283
0.0446
0.3094
0.3595
0.3762
0.4573
F2
0.0000
0.1969
0.0143
0.0127
0.0116
0.0115
0.0069
0.2178
0.2191
0.2236
0.2257
0.0201
0.0494
0.3539
0.4083
0.4277
0.5129
CONCLUSIONS AND FUTURE WORK
In this work, we proposed an approach for automatically
detecting information supplied in confidence in government
documents. Our approach uses part-of-speech n-grams to
measure the sensitivity load of sequences of text. On a
collection of sensitivity-reviewed government documents, we
showed that the sensitivity load of these sequences can be
used to accurately classify in confidence sensitivities within
government documents. In particular, using a CRF sequence
tagger with the sensitivity load of POS n-grams as features,
this approach achieved over 0.45% recall and 0.99% precision, markedly outperforming a baseline approach from the
literature that has been shown to achieve high levels of recall for sensitive text in other domains. As future work, we
intend to conduct a user study to quantify the benefits of
our approach for assisting in the sensitivity review of digital
government documents.
10.
ACKNOWLEDGMENTS
The authors would like to thank the sensitivity reviewers
for their help in constructing the test collection.
11.
REFERENCES
[1] D. Abril, G. Navarro-Arribas, and V. Torra. On the
declassification of confidential documents. In Mod. Dec. for Art.
Intel.. 6820:235–246, 2011.
[2] D. A. R. P. Agency. Darpa, new technologies to support
declassification. Request for Information, DARPA-SN-10-73, 2010.
[3] A. Allan. Records Review. UK Government. 2014.
[4] J. R. Finkel, T. Grenager, and C. Manning. Incorporating
non-local information into information extraction systems by
gibbs sampling. In Proc. of ACL, 2005.
[5] T. Gollins, G. McDonald, C. Macdonald, and I. Ounis. On using
information retrieval for the selection and sensitivity review of
digital public records. In Proc. of PIR at CIKM, 2014.
[6] J. D. Lafferty, A. McCallum, and F. C. N. Pereira. Conditional
random fields: Probabilistic models for segmenting and labeling
sequence data. In Proc. of ICML, 2001.
[7] Y. Li, C. Luo, and S. Chung. Text clustering with feature
selection by using statistical data. In Trans. on Knowledge and
Data Engineering, 20(5):641–652, 2008.
[8] C. Lioma and I. Ounis. Examining the Content Load of
Part-of-Speech Blocks for Information Retrieval. In Proc. of
COLING/ACL, 2006.
[9] G. McDonald, C. Macdonald, I. Ounis, and T. Gollins. Towards
a classifier for digital sensitivity review. In Proc. of ECIR, 2014.
[10] T. Mendel. Freedom of information: a comparative legal survey.
Paris: Unesco, 2008.
[11] D. Sánchez, M. Batet, and A. Viejo. Detecting sensitive
information from textual documents: an information-theoretic
approach. In Mod. Dec. for Art. Intel., 7647:173-184, 2012.